A 51% leap in cryosphere forecasting: How a new AI-powered method is rewriting climate history

The cryosphere—the planet’s vast store of snow, ice, and frozen ground—is vanishing at an accelerating pace. Its fate is central to sea level rise, water security, and global climate stability. Yet for decades, scientists have struggled to reconstruct its past with confidence, especially in remote regions where data are sparse and models uncertain. Now, a breakthrough method has emerged that improves the accuracy of cryospheric reanalysis by 51%—a leap that could transform how we understand and predict the behavior of Earth’s frozen zones.

This isn’t just a statistical tweak. It’s a fundamental rethinking of how we blend climate models with real-world observations. The new approach, called climate-informed cryospheric reanalysis, uses a hierarchical Bayesian framework to strike a balance between two long-standing extremes: treating each year in isolation, or assuming every year follows the same unchanging pattern. By doing so, it captures both the unique quirks of individual years and the deeper rhythms of regional climate—seeing, in effect, both the trees and the forest.

And it does so at a fraction of the computational cost of previous gold-standard methods, opening the door to high-resolution, uncertainty-aware reconstructions of snowpacks and glaciers across the globe.

The Science

Cryospheric reanalysis is the process of reconstructing the historical state of snow, ice, and permafrost by combining physical models with observations—much like weather reanalysis does for the atmosphere. The goal is to produce a continuous, physically consistent record of variables like snow water equivalent (SWE) or glacier mass balance, even in years with few or no direct measurements.

Traditionally, two strategies have dominated:

1. No pooling: Each water year is treated independently. Model parameters—like how efficiently snow accumulates or melts—are inferred separately for each year using only that year’s data. This captures interannual variability but fails when data are sparse, causing the model to revert to poorly calibrated assumptions.

2. Complete pooling: All years are lumped together to estimate a single set of static parameters. This stabilizes estimates in data-poor years but erases year-to-year differences, making the model blind to climate extremes.

The new method, developed by Aalstad et al. (2026), introduces a third way: partial pooling. It uses a hierarchical Bayesian model where annual parameters are drawn from a shared climatological distribution—defined by hyperparameters that are themselves inferred from the data. This allows information to flow across years, but not uniformly: in data-rich years, the model trusts the observations; in data-sparse years, it leans on the inferred climate baseline.

The framework is implemented through a nested hybrid particle smoothing workflow (

), combining expectation maximization (EM) with a particle-based smoother called MAGPIES and an adaptive batch smoother (AdaPBS). This hybrid design avoids the computational burden of full Markov chain Monte Carlo (MCMC) sampling while maintaining high accuracy.

The team tested the method at eight sites across Europe and the Arctic, from the Swiss Alps to Svalbard (

), using a temperature index model—a widely used, computationally efficient tool for simulating snow and glacier melt. Three experiments were conducted:

• In situ SWE assimilation at four snow sites, using direct snow measurements.
• Satellite-based FSCA assimilation using MODIS data on snow-covered area.
• Glacier mass balance assimilation at four glaciers, using long-term glaciological records.

Each experiment compared the performance of no pooling, complete pooling, and the new partial pooling approach during calibration (1941–1999) and validation (2000–2024).

What They Found

The results were striking. Across all experiments, the partial pooling method outperformed both alternatives in nearly every metric.

The most comprehensive measure was the continuous ranked probability score (CRPS), which evaluates the accuracy and sharpness of probabilistic forecasts. Relative to the prior (the model’s initial guess), partial pooling improved CRPS by 51%—far exceeding the 38% improvement from no pooling and the 19% from complete pooling.

CRPS Improvement Relative to Prior

Label	Value
Partial Pooling	51 %
No Pooling	38 %
Complete Pooling	19 %

Partial pooling yields a 51% improvement in continuous ranked probability score, far exceeding no pooling (38%) and complete pooling (19%).

This wasn’t just a win on paper. At Filefjell, Norway, where in situ SWE data were assimilated, the partial pooling reanalysis captured both the long-term decline in snowpack and the sharp interannual swings driven by weather extremes (

). In contrast, the no pooling approach produced erratic estimates in years with missing data, while complete pooling smoothed over real variability.

Similarly, at Hornsund in Svalbard, where only satellite-based fractional snow-covered area (FSCA) was available, the climate-informed method reconstructed SWE with remarkable fidelity—even when validated against independent in situ measurements (

). The no pooling model again struggled in sparse years, while complete pooling failed to adapt to changing conditions.

For glaciers, the gains were equally clear. When estimating annual mass balance, partial pooling reduced uncertainty and improved correlation with observations. In the validation period, it correctly captured the accelerating mass loss at Claridenfirn in the Swiss Alps and Austre Brøggerbreen in Svalbard—trends that complete pooling underestimated and no pooling often missed entirely (

).

Glacier Mass Balance Reconstruction

Label	Value
1940	-0.2
1950	-0.25
1960	-0.3
1970	-0.35
1980	-0.4
1990	-0.5
2000	-0.6
2010	-0.7

Reconstructed glacier mass balance (blue) closely tracks observed trends (orange), capturing accelerating ice loss from 1940–2020.

The method also proved computationally efficient. Compared to a particle MCMC benchmark, the hybrid workflow reduced computational cost by several orders of magnitude while achieving comparable or better skill. This efficiency is critical for scaling the approach to regional or global applications.

Why This Changes Things

The implications of this work extend far beyond academic metrics. For the first time, we have a scalable, uncertainty-aware method to reconstruct the cryosphere’s past with high fidelity—even in data-poor regions.

Consider water security. Over 1.9 billion people rely on seasonal snowmelt for drinking water, agriculture, and hydropower (Barnett et al., 2005). In the western US, the Andes, and High Mountain Asia, snowpack acts as a natural reservoir, storing winter precipitation and releasing it slowly in spring and summer. But as climate change alters snowfall patterns, knowing how much water is stored—and how that has changed over time—is essential for managing reservoirs, droughts, and floods.

Current snow reanalyses often rely on assumptions that degrade accuracy in remote or complex terrain. The new method, by contrast, can leverage sparse in situ data and satellite observations to produce reliable, probabilistic estimates of SWE across decades. This could transform water resource planning in regions like the Himalayas, where glacial melt feeds the Indus, Ganges, and Brahmaputra rivers.

Then there’s sea level rise. Glaciers outside Greenland and Antarctica are now the largest contributors to global sea level rise, surpassing even the ice sheets in some years (Hugonnet et al., 2021). To project future rise, we need accurate estimates of past mass loss—and the drivers behind it. The partial pooling approach allows us to infer long-term glacier behavior even when direct observations are intermittent, improving the reliability of ice loss models.

Beyond practical applications, this work represents a conceptual shift in how we do Earth system science. For too long, the field has been split between two worlds: highly detailed but computationally expensive models, and fast but simplistic ones. Between data-rich and data-poor regions. Between deterministic forecasts and probabilistic uncertainty.

The climate-informed reanalysis framework bridges these divides. It shows that we don’t have to choose between resolution and uncertainty, or between local detail and global coherence. By embedding climate knowledge directly into the inference process, it creates a feedback loop where data improve the model, and the model improves the interpretation of data.

This is especially valuable in the Arctic, where warming is amplifying three to four times faster than the global average (Meredith et al., 2019). There, sparse observations and extreme conditions have long made modeling a challenge. Yet the stakes are high: thawing permafrost could release vast amounts of carbon, and shrinking sea ice is reshaping global weather patterns.

The partial pooling method offers a way to extract maximum information from limited Arctic data, helping us track changes in snow cover duration, glacier retreat, and permafrost stability with unprecedented clarity.

What’s Next

The success of this method opens several promising paths forward.

First, scaling up. The current experiments were run at point locations, but the workflow is modular and could be extended to gridded, regional, or even global reanalyses. Combined with emerging satellite data—like NASA’s upcoming NISAR mission, which will provide high-resolution snow depth and glacier velocity maps—this could enable the first high-fidelity, uncertainty-quantified cryospheric reanalysis spanning decades and continents.

Second, model integration. While the study used a simple temperature index model, the framework is model-agnostic. It could be applied to more complex models like CryoGrid (Westermann et al., 2023) or the Community Land Model, allowing for richer process representation while still benefiting from hierarchical inference.

Third, real-time applications. The computational efficiency of the hybrid smoother makes it suitable not just for reanalysis, but for near-real-time monitoring. Imagine a system that updates daily, blending satellite snow cover data with ground measurements to produce probabilistic forecasts of snowmelt runoff—critical for flood warning and reservoir management.

Finally, machine learning synergy. The hierarchical Bayesian approach shares deep conceptual roots with modern AI. In fact, deep neural networks can be seen as approximations of Bayesian models—albeit ones that often lack uncertainty quantification (Murphy, 2023). By combining the physical constraints of cryospheric models with the flexibility of machine learning, future systems could use this framework to learn not just parameters, but entire model structures from data.

Challenges remain. The method assumes that the climatological hyperparameters are stationary over time—a reasonable assumption for now, but one that may break down under rapid climate change. Future versions could incorporate time-varying hyperparameters to capture shifting baselines.

There’s also the question of observational bias. Satellite retrievals of snow cover or glacier elevation can have systematic errors, especially in complex terrain. The current framework treats observation errors as known, but a fully hierarchical model could also infer error characteristics from the data.

Still, the core insight—that we can learn from both individual events and long-term patterns—is powerful. It reflects a broader shift in science toward integrative, uncertainty-aware modeling—one that doesn’t just predict the future, but quantifies how sure we can be.

As the cryosphere continues to transform, we need more than snapshots. We need narratives—coherent, probabilistic stories of how snow and ice have changed, why, and what comes next. This method brings us closer to that goal.

The Earth’s frozen zones are not just remote landscapes. They are climate regulators, water towers, and cultural landmarks. By giving us a clearer view of their past, this work helps us prepare for their uncertain future.